gengar-atropos-environments

What problem does it solve?

It helps you implement and troubleshoot Gengar reinforcement learning environments for Atropos training without breaking the agent loop, tool calling, or reward evaluation contracts.

Core Features & Use Cases

• GengarBaseEnv integration guide: Implements the required lifecycle (setup, get_next_item, format_prompt, compute_reward, evaluate, wandb_log) so the environment works with the Gengar multi-turn agent loop.
• Correct reward and evaluation wiring: Ensures compute_reward scores using AgentResult.messages and ToolContext sandbox verification, while evaluate runs the full GengarLoop (tools included).
• Production-friendly CLI workflows: Covers the three modes—serve, process, and evaluate—with provider-agnostic inference setup prompts and correct flags.

Quick Start

Ask the AI to walk you through creating your environments/your_env.py by mapping each required method to the correct GengarBaseEnv interfaces, including how to score rollouts with compute_reward and verify them via ToolContext in evaluate mode.